Paste a benchmark claim. See what it forgot to say.
A 30-second forensic breakdown of any ML accuracy, refusal-rate, or pass-rate claim. We check it against 9 falsifiability criteria the spec considers minimum hygiene. Heuristic, not authoritative, but it surfaces what most published claims quietly omit.
This is a heuristic regex-based check, not a formal PRML verifier. Some honest claims will score low because they describe their methodology elsewhere; some dishonest ones might gain points by mentioning the right keywords without backing them up. The score surfaces structural omissions. Interpret accordingly. A high score means the claim is checkable, not that it is true: A PRML receipt shows the bar was on record by a given time, never that the result is right. The real spec is at spec.falsify.dev/v0.1.
If this surfaced something you didn't expect
The 9 criteria above aren't arbitrary: they're the structural fields PRML asks you to commit to a SHA-256 hash before you run an evaluation. Once committed, the receipt shows the threshold and metric were on record no later than the receipt time; it does not show when the evaluation ran.